PFENet++: Boosting Few-Shot Semantic Segmentation With the Noise-Filtered Context-Aware Prior Mask.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Prior mask guidance improves few-shot segmentation by highlighting regions of interest.
• Current methods often overlook contextual information in prior mask generation.
• Maximum correlation can be sensitive to noisy features, impacting mask quality.

Purpose of the Study:

• To enhance prior mask generation for few-shot semantic segmentation.
• To improve object localization using contextual semantic cues.
• To develop a robust method less susceptible to noisy features.

Main Methods:

• Proposed Context-aware Prior Mask (CAPM) leveraging nearby semantic cues.
• Introduced a lightweight Noise Suppression Module (NSM) to filter noisy features.
• Integrated CAPM and NSM into the PFENet framework, creating PFENet++.

Main Results:

• PFENet++ significantly outperforms the baseline PFENet and other competitors.
• Achieved state-of-the-art results on PASCAL-5^i, COCO-20^i, and FSS-1000 benchmarks.
• Demonstrated substantial practical merit and maintained efficiency.

Conclusions:

• The proposed CAPM and NSM effectively improve prior mask quality and object localization.
• PFENet++ establishes a new state-of-the-art in few-shot semantic segmentation.
• The model shows potential as a strong and efficient baseline for future research.